# Radiomic analysis of contrast-enhanced CT for the prediction of microvascular invasion in hepatocellular carcinoma: literature analysis and practical challenges

## TL;DR

This study explores using radiomic features from CT scans to predict microvascular invasion in liver cancer, but finds that manual segmentation limits reproducibility.

## Contribution

The study highlights the practical challenge of manual segmentation affecting reproducibility in radiomic analysis for predicting MVI in HCC.

## Key findings

- A bivariate model using two radiomic features achieved 79% AUC for predicting MVI.
- Segmentation robustness was strongly influenced by reader experience.
- Inter-reader reproducibility was suboptimal for less experienced operators.

## Abstract

Microvascular invasion (MVI) is considered an independent risk factor for early recurrence after curative resection of hepatocellular carcinoma (HCC). The ability to preoperatively predict MVI could lead to personalized treatment options in high-risk patients.

To identify radiomic features from CE-CT that correlate with MVI in patients with HCC and evaluate the robustness and reproducibility of radiomic assessment by manual segmentation between readers with different experience.

Clinical, CT imaging, and histological parameters were recorded. Sixty-two HCC lesions were manually contoured by three radiologists. Radiomic features were extracted. Features best correlating with angioinvasion were selected and assessed in univariate and multivariate models by means of 100 trials of 5-fold stratified cross-validation in terms of AUC, sensitivity, and specificity. The model identified on contours from the most experienced operator was then tested on contours from the other operators to assess inter-reader reproducibility.

Feature selection identified LI-RADS category and four arterial-phase radiomic texture features, with GLCM-ClusterShade and its high-frequency wavelet variant showing the highest predictive value for MVI. A bivariate logistic regression model combining these two features achieved an AUC of 79%, with 78% sensitivity and 64% specificity. The robustness of manual segmentation was strongly dependent on reader experience, and inter-operator reproducibility was suboptimal when the model was applied to contours from less experienced readers.

Radiomics analysis may be able to predict MVI in patients with HCC. However, segmentation methods remain a practical challenge affecting reproducibility in radiomic studies.

This study, in agreement with the literature, identifies a radiomic model based on two textural features that could correlate with MVI in HCC. Furthermore, it aims to investigate some of the limitations in the application of radiomics in clinical practice, which still restrict it to a research setting, identifying an important limitation in manual segmentation methods. This aspect has not yet been sufficiently investigated in the literature.